Leveraging the Potential of Machine-Learning Interatomic Potentials for QM/MM Simulations

Area of Science:

• Computational Chemistry
• Materials Science
• Biophysics

Background:

• Machine-learning interatomic potentials (MLIPs) are emerging as efficient alternatives to quantum-mechanical (QM) calculations for molecular dynamics (MD) simulations.
• MLIPs offer accuracy comparable to QM methods like density functional theory (DFT) when trained on sufficient data.
• Despite their advantages, MLIPs remain computationally intensive for large biological systems, limiting simulation timescales and system sizes.

Purpose of the Study:

• To review recent advancements and current developments in multiscale ML/MM approaches for simulating large biological systems.
• To highlight ML/MM as a viable compromise between computational cost and simulation scale for complex systems in solution.

Main Methods:

• Overview of MLIPs and their integration into multiscale simulation frameworks.
• Discussion of ML/MM methodologies analogous to QM/MM.
• Analysis of computational trade-offs between MLIPs, QM, and classical force fields (MM).

Main Results:

• MLIPs provide a significant speedup over QM methods while maintaining high accuracy.
• Multiscale ML/MM methods enable the simulation of larger biological systems and longer timescales compared to full MLIP simulations.
• Current developments focus on optimizing ML/MM performance and applicability to diverse biological problems.

Conclusions:

• ML/MM approaches represent a promising strategy for bridging the gap between computational cost and simulation capabilities in complex biological systems.
• Continued research in ML/MM is crucial for advancing molecular simulations in biophysics and computational chemistry.
• These methods facilitate more comprehensive investigations of biological processes at the molecular level.